My Community

Hi,

I am new to VASP and trying to understand memory usage. The OUTCAR prints the "total amount of memory used by VASP MPI-rank0." I believe this is the memory used by a single core in my nodes.

I have 640 atoms and more than a million plane waves in my system, and I'm just trying to simulate a simple QMD run in an NPT ensemble. Since I have a lot of plane waves, I used 32 nodes, which consist of 56 cores per node and 190GB of RAM for each node. This leads to 56*32 = 1792 total tasks. I get an "insufficient virtual memory" from Slurm, which I believe is an insufficient memory issue.

But when I calculate the total memory requirement per node based on the "total amount of memory used by VASP MPI-rank0 44223. kBytes", it should be about 2.8 GB per node (44223 * 56 kBytes), which is much less than the total RAM of 190 GB/node.

Can someone help me understand this? Appriciate it verymuch.

Best,
Chamara

Hello,

Memory usage is a challenging topic, but with such large systems it is an important consideration.

In the OUTCAR you can find

total amount of memory used by VASP MPI-rank0

but that is just for rank0 and cannot be simply multiplied by the number of ranks.
There is more information below

 General timing and accounting informations for this job:

Still, these numbers will not agree with the actual use of memory. That could only be tracked from the outside via tools like top, memcheck etc.

Maybe more important than the exact memory usage, ist the distribution and scaling with different settings. I recommend starting with a smaller system (smaller supercell) and looking at how the demands increase when increasing the supercell.

By the way, with QMD run you are referring to ab initio molecular dynamics, right? I think NCORE is an important setting for you to look into and find the optimal setting.

Best regards,
Marie-Therese

Hello!

Just a small addition to Marie-Therese's answer... maybe the message "insufficient virtual memory" hints at an upper limit that was set for this HPC system. You could try to look at

ulimit -v

to check the upper limit on virtual memory which is allowed on this system. It could return a number in kB or the word "unlimited". Btw, "ulimit -a" gives you all limits, maybe you could share the output here?

The usage of virtual memory you can monitor also with the command "top" but be aware that this has nothing to do with the actual amount of RAM required. So maybe there is actually enough RAM memory installed to run your simulation but unfortunately the virtual memory limit prevents its execution. One can unset the limit with

ulimit -v unlimited

but I am afraid that this can usually only be done by a system administrator.

Anyway, I also suggest to start with a less memory-demanding setup first and try to scale up slowly, so you can extrapolate how much memory the full simulation would require.

All the best,
Andreas Singraber